Local cache for AI agents
Stop paying
twice
for AI.
OmniCache answers repeat calls from Claude Code, Cursor and any LLM app on your own machine, in milliseconds, without spending a token.
Runs on localhost No code changes Python 3.9+
Your agent
OmniCache
LLM API
Same session.
A fraction of the cost.
Wait time−75%
Tokens billed−75%
Illustrative 4-turn session re-reading one unchanged file: one miss at 1,450 ms and 820 tokens, then three local hits at 1.8 ms each.
Measured on a 4-vCPU machine
9 µsexact-match lookup (p50)
0.18 msnear-duplicate lookup (p50)
0 / 20false matches on unrelated prompts
Benchmark report
Four checks.
Under a millisecond.
-
01
Intercept
Sits on localhost between your agent and the API.
-
02
Verify
Git SHA and mtime checks. Edited files are never replayed.
-
03
Replay
Exact repeats are served from memory.
-
04
Match
Reworded prompts match too, with stricter rules for code.
Miss? Forwarded once, then cached.
One install.
Zero config.
$ pip install omnicache-proxy$ omnicache run claude✓ Proxy on http://127.0.0.1:8000✓ ANTHROPIC_BASE_URL set · launching claudeWorks with Claude Code · Cursor · Cline · Aider · any OpenAI- or Anthropic-compatible client